Accurate and rapid identification of corn leaf diseases is crucial for sustainable agriculture and healthy crops. Traditional CNN-based disease classification models have demonstrated decent performance in controlled settings but often fail to generalize well across real-word scenarios. In this study, we propose DinoCorn, a robust vision-based disease classification model leveraging the DINOv2 vision transformer as a frozen feature with a custom-trained classification head for precise classification of corn leaf diseases. We introduce a newly consolidated dataset with 6516 expert rated images from various sources across four corn leaf classes (Gray Leaf Spot, Common Rust, and Northern Leaf Blight, Healthy) - collected from PlantVillage and Seasonal Corn Leaf Disease Dataset for training, validation and testing DinoCorn. Our results demonstrate DinoCorn’s superior performance compared to pre-trained CNN models such as Resnet50, Inceptionv3, and Mobilenetv3. DinoCorn achieved an accuracy of 97.8% with an error rate of 2.2% and an F1-score of 0.98, significantly outperforming accuracy of ResNet50 (91.9%), Inceptionv3 (91.4%), and MobileNetv3 (88.3%) and F1-score of ResNet50 (0.92), Inceptionv3 (0.91), and MobileNetv3 (0.88). The results indicate strong generalization across varied lighting, background and leaf conditions, making DinoCorn a promising solution for field-ready deployment in plant disease classification systems. Furthermore, DinoCorn’s low inference latency of 0.138 ms suggests its suitability for real-time leaf disease monitoring systems, enabling timely interventions. These findings underscore the potential of advanced vision transformers in revolutionizing corn leaf disease prediction and management in modern agriculture.
Kaggle Link: https://www.kaggle.com/datasets/yasirahmad0810/consolidated-corn-dataset
Kaggle Notebook: https://www.kaggle.com/code/ajaynidadavolu/dinov2-corn2-final